Molecular generative models learn the grammar of chemistry from large compound libraries, but when we ask them to work in an unfamiliar region of chemical space — natural products instead of synthetic screening compounds, say — they fail in two characteristic ways. Fine-tuned directly on the small target set, they forget the broad grammar they learned (the classic catastrophic-forgetting problem); adapted gently, they drift back to the handful of core structures they already knew, a failure mode that is invisible to standard loss curves because the molecules remain perfectly valid. We propose to fix both problems with one piece of machinery: an adversarial critic that, instead of issuing a single real-or-fake score, sorts every molecule into one of $K$ chemical-structure families or a "this is off the real manifold" bucket. Such a critic gives the generator a class-aware gradient signal, and it gives us, for free, a live measurement of how much of chemical space the model is actually covering.
The specific setting is sequential domain adaptation of a molecular generator: a staged curriculum that moves a pre-trained model from broad chemical space through progressively more similar intermediate regions to a small, structurally biased target. Prior work tells us the order of these stages matters — adaptation must run from the domain least similar to the target to the one most similar, or the target distribution is destroyed One-Shot Learning for Multi-Domain Neural Machine Translation — but also that no principled method exists for measuring that similarity; ordering is done by trial and error or crude text-level scores. Our claim is that a categorical discriminator over scaffold clusters solves the ordering problem and the coverage problem simultaneously, and that this combination — categorical adversarial feedback plus a derived curriculum — is what current molecular domain-adaptation methods leave on the table.
Core Claim
The proposal transplants the $(K+1)$-class adversarial discriminator — a critic whose output is a softmax over $K$ real categories plus one synthetic category, which stabilizes class-conditional generation in image GANs Activation Maximization GANs and multi-site image normalization Realistic image normalization for multi-Domain segmentation — into molecular generation, where the $K$ categories are scaffold clusters rather than image classes. Concretely, we partition chemical space by extracting Bemis–Murcko scaffolds and clustering them, the same stratification that MoleculeNet standardized to expose models that memorize local motifs MoleculeNet: A Benchmark for Molecular Machine Learning. A discriminator $D$ maps a molecule to a probability over the $K$ scaffold clusters plus one off-manifold class, trained on real molecules labeled by cluster and on negatives that are chemically valid but distributionally foreign:
$$ \mathcal{L}_D = -\,\mathbb{E}_{x,\,y}\big[\log D_y(x)\big], \qquad y \in \{c_1, \dots, c_K, \text{ off-manifold}\}. $$
This single trained object does three jobs that no existing molecular pipeline assigns to one component. First, it orders the curriculum: the average class-logit distribution of each candidate intermediate domain defines a distance to the target domain in the discriminator's representation space, replacing the empirical trial-and-error that currently plagues furthest-to-closest ordering. Second, it steers adaptation: during each stage the generator is rewarded for producing molecules the discriminator places in the current stage's real scaffold clusters, receiving a class-aware gradient vector rather than a scalar — the same mechanism that lets AM-GAN counteract the overlaid-gradient problem, where gradients from multiple classes cancel and pull the generator into blurry non-class territory Activation Maximization GANs. Third, it measures coverage: the Shannon entropy of the batch-average scaffold-class distribution,
$$ C = H\!\left(\tfrac{1}{B}\sum_{i=1}^{B} D_{1:K}(x_i)\right), $$
falls when a generator collapses into a few familiar scaffolds, giving a live coverage meter tied directly to the chemical-space-coverage problem rather than to proxy loss values.
Research Program
We build a SELFIES-based sequence-to-sequence generator — SELFIES strings guarantee 100% chemically valid molecules by construction, as exploited by MOLGEN Domain-Agnostic Molecular Generation with Chemical Feedback — pre-trained on a broad corpus (GuacaMol, following the molecular-adaptation work of Sultan et al. Transformers for molecular property prediction: Domain adaptation efficiently improves performance). On top of this we train the scaffold-class discriminator, derive the curriculum, and run staged adaptation.
The adaptation loop at each stage combines three signals: the categorical adversarial objective above, dynamic labeling in the style of AM-GAN — the target cluster for each generated sample is reassigned each step to the cluster where the discriminator currently scores it highest, which prevents intra-cluster mode collapse that rigid label assignment causes Activation Maximization GANs — and replay of earlier-stage molecules to resist forgetting, motivated by the finding that reversing stage order destroys the target distribution in multi-domain translation One-Shot Learning for Multi-Domain Neural Machine Translation. The coverage meter $C$ gates stage transitions: the model advances only when its generated distribution's scaffold-class entropy saturates, and triggers replay when entropy collapses mid-stage.
Two design risks need explicit guards, both flagged by prior work. The first is adversarial exploitation: AM-GAN's authors showed generators can trigger high categorical logits with adversarial noise that lacks valid semantics Activation Maximization GANs. SELFIES removes the gross form of this (invalid valence), but a molecule could still satisfy the discriminator while being chemically absurd; we therefore evaluate all coverage claims with an independent scaffold classifier held out from adversarial training, and filter generations with RDKit physicochemical sanity checks. The second is hyperparameter fragility of composite adversarial objectives, documented for residual image translation Learning Residual Images for Face Attribute Manipulation; we fix a single loss-weighting recipe across all domain pairs and report sensitivity, rather than hand-tuning per pair.
The main testbed is cross-domain generation from synthetic compounds (ZINC-15) to natural products (NPASS), the same two-domain pair as MOLGEN, but traversed sequentially through intermediate regions (DrugBank compounds and GEOM-derived scaffold clusters, resources already used for transfer evaluation in MolPLA's core/R-group framework MolPLA: A Molecular Pretraining Framework for Learning Cores, R-Groups and their Linker Joints). A second testbed adapts toward scaffold clusters held out from MoleculeNet benchmarks (BBBP, BACE, ESOL, FreeSolv) so that downstream property prediction can be scored under the standard scaffold-split protocol MoleculeNet: A Benchmark for Molecular Machine Learning.
Baselines
The naive reference point is the un-adapted pre-trained generator asked to produce target-domain molecules, plus uniformly sampled SELFIES strings as a floor for what "off-manifold" means. This measures raw cross-domain transfer with no machinery at all; any method worth its cost must at least widen the gap between these two reference points.
The strongest existing competitor is MOLGEN's domain-agnostic prefix tuning, which adapts a BART generator between ZINC and NPASS using soft domain-instructor prefixes and scalar property-ranking feedback Domain-Agnostic Molecular Generation with Chemical Feedback. It is the closest published system on exactly our testbed, and it reports a penalized logP optimization score of 80.30 (versus 44.99 for MARS) with 100% validity. We will re-run it under matched conditions: same SELFIES backbone, same pre-training corpus, same compute and tuning budget, same intermediate-domain data. Beating it means achieving higher target-scaffold coverage — measured by the independent held-out classifier as the fraction of generated molecules landing in target-region clusters, including clusters under-represented in the target training set — at equal validity, and matching or exceeding its reported 80.30 on penalized logP. If prefix tuning matches our coverage at lower cost, the adversarial machinery is not justified.
The ablation that isolates our key ingredient is a staged-adaptation pipeline identical in every respect — same curriculum order, same data, same replay schedule — but with the discriminator replaced by non-adversarial alignment objectives: masked-language-model and multi-task physicochemical-descriptor regression, the two objectives that Sultan et al. found most effective for intermediate molecular domain adaptation, where the descriptor-based version added roughly 0.2 in $R^2$ over generic masked-language adaptation Transformers for molecular property prediction: Domain adaptation efficiently improves performance. This isolates whether the categorical adversarial signal specifically, rather than staged adaptation per se, drives any gains. Two further ablations decompose the discriminator's two roles: a binary real/fake discriminator per stage (removing categorical structure while keeping adversarial feedback), and a hand-fixed or Tanimoto-distance-ordered curriculum (removing the learned proximity signal while keeping categorical feedback).
Finally, a joint-training baseline pools all domains with the $(K+1)$ discriminator active simultaneously, in the spirit of Delisle et al.'s multi-site medical normalization, where the multi-class discriminator aligns all cohorts in one game rather than a sequence Realistic image normalization for multi-Domain segmentation. This tests whether sequencing matters at all once the categorical critic is present.
Why not just…?
Why not just use prefix tuning? This is the most serious alternative, and we treat it as the primary baseline rather than arguing around it. MOLGEN demonstrates that soft prefixes move a generator between two hand-picked domains effectively Domain-Agnostic Molecular Generation with Chemical Feedback. But the domain boundary is left implicit, the feedback is a scalar ranking signal optimized for single endpoints, and nothing in the method detects coverage collapse or scales beyond two domains chosen by hand. Our design bets that class-conditional gradients plus an explicit proximity measurement generalize to many domains and small, biased targets. If the head-to-head shows prefix tuning matching our coverage at the same budget, that is a decisive negative result for the adversarial component, and we would report it as such.
Why not just order the curriculum with a cheap structural distance, like fingerprint Tanimoto similarity? That is the obvious way to get a proximity metric without training anything, and the file's own survey of sequential adaptation notes that ordering currently relies on simple proxy measures because no representation-space metric exists One-Shot Learning for Multi-Domain Neural Machine Translation. The counter-evidence is that structural similarity and functional similarity dissociate: in crystallographic fragment screening, compounds with near-zero structural similarity showed identical protein-interaction fingerprints, and structurally diverse libraries were functionally redundant Fragment libraries designed to be functionally diverse recover protein binding information more efficiently than standard structurally diverse libraries. The discriminator's ordering is trained against the very representation the generator uses, so it should track transfer-relevant proximity better than Tanimoto distance does. This is directly testable: the ordering ablation compares curricula derived from discriminator logits, Tanimoto distance, and a deliberately reversed order. If the Tanimoto ordering performs as well, the discriminator's ordering role is unnecessary — though its feedback and coverage roles would remain untested, and we would say so.
Why not just pool all domains and train jointly, or fine-tune directly on the target? Joint pooling has real empirical support: in multi-domain machine translation, training on all concatenated data followed by target fine-tuning was the best configuration, at 35.1 BLEU (Arabic–English) and 35.5 (German–English) One-Shot Learning for Multi-Domain Neural Machine Translation, and Delisle et al.'s joint multi-domain game works in medical imaging Realistic image normalization for multi-Domain segmentation. But those settings have large, roughly comparable domains; our motivation is the regime where the target is tiny and structurally biased, so pooling washes out rare scaffolds — the signal-dilution problem that motivates staged adaptation in the first place Domain-Adaptive Pretraining — and direct fine-tuning on such a target is exactly what destroys the base representation. The joint-game and direct-fine-tuning baselines settle this empirically: if joint pooling with the categorical critic matches the sequential curriculum, the sequencing claim fails and the paper's contribution narrows honestly to the critic itself.
Proposed Experiments and Decisive Checks
The first experiment measures cross-domain coverage on ZINC → NPASS through intermediates. Metrics are the fraction of generated molecules assigned by the independent scaffold classifier to target-region clusters, the number of distinct target clusters reached (coverage breadth), internal diversity, and novelty against the target training set. The decisive comparison is against prefix tuning under matched budgets; the secondary comparison is against the non-adversarial staged ablation, which tells us whether categorical adversarial feedback beats the best non-adversarial alignment objectives that Sultan et al. identified Transformers for molecular property prediction: Domain adaptation efficiently improves performance.
The second experiment validates the coverage meter itself: does the discriminator's batch entropy $C$ track externally measured scaffold-class coverage across training, and does it flag collapse earlier than loss curves or validity rates? This is a measurement-methods contribution that stands even if the generation gains are modest, because no current molecular adaptation method has any live coverage signal.
The third experiment tests curriculum ordering: discriminator-derived order versus Tanimoto order versus reversed order, all with the categorical critic active. The hypothesis, drawn from the translation literature One-Shot Learning for Multi-Domain Neural Machine Translation, is that reversed order degrades target coverage sharply while the two sensible orders differ less — and that the discriminator order beats Tanimoto when structural and functional proximity disagree, as the fragment-screening evidence predicts Fragment libraries designed to be functionally diverse recover protein binding information more efficiently than standard structurally diverse libraries.
The fourth experiment scores downstream utility on scaffold-split MoleculeNet benchmarks (BBBP, BACE, ESOL, FreeSolv) and penalized logP optimization against MOLGEN's reported 80.30 Domain-Agnostic Molecular Generation with Chemical Feedback. For context, the best hierarchical graph model reports ROC-AUC of 0.954 on BBBP and 0.890 on BACE under scaffold splits Learning Hierarchical Interaction for Accurate Molecular Property Prediction; no reported result found on Lacuna for a generative-adaptation method on these benchmarks, which is itself part of the gap. We hold this expectation honestly low: Sultan et al. found that even their best adapted transformer lost to a random forest on raw 2D descriptors on 5 of 7 ADME datasets Transformers for molecular property prediction: Domain adaptation efficiently improves performance, so we treat property prediction as a secondary check that adapted generations are chemically useful, not as the headline claim.
The decisive check for the whole proposal is simple to state. If, under matched data and compute, the scaffold-class discriminator's curriculum reaches more target-region scaffold clusters at equal validity than prefix tuning, non-adversarial staged adaptation, and the joint game — and if its batch entropy predicts that coverage — then categorical adversarial feedback is the right instrument for moving molecular generators across chemical space, and the ordering and coverage signals it provides are free byproducts. If it does not, the honest conclusion is that soft prompting and non-adversarial objectives suffice for two-domain transfer, and the interesting open question becomes whether any method can detect coverage collapse online at all.
References
- Neural Machine Translation Training in a Multi-Domain Scenario
- Activation Maximization Generative Adversarial Nets
- Realistic Image Normalization for Multi-Domain Segmentation
- MoleculeNet: a benchmark for molecular machine learning
- Domain-Agnostic Molecular Generation with Chemical Feedback
- Transformers for molecular property prediction: Domain adaptation efficiently improves performance
- Learning Residual Images for Face Attribute Manipulation
- MolPLA: a molecular pretraining framework for learning cores, R-groups and their linker joints
- Fragment Libraries Designed to Be Functionally Diverse Recover Protein Binding Information More Efficiently Than Standard Structurally Diverse Libraries
- Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
- Learning Hierarchical Interaction for Accurate Molecular Property Prediction